import numpy as np
import matplotlib.pyplot as pl
from typing import Callable

float_formatter = "{:.3f}".format
np.set_printoptions(formatter={'float_kind': float_formatter})

K = 18


def bandit(a: int, bias: np.array):
	zn = np.random.uniform() + bias[a]
	return zn


def test(_episode_len: int, _qu: np.array, _nu: np.array, _bandit: Callable) -> np.array:
	_epsilon = 0.5		# Anti-greediness
	_rho = 0.005		# reduce epsilon with age
	_r_sum = 0
	_r_mean = []
	_n = 0
	for j in range(0, _episode_len):
		# choose action
		z = np.random.uniform()
		if z <= _epsilon:
			# Choose random action
			a = np.random.randint(low=0, high=K)
		else:
			# Choose best action
			a = np.argmax(_qu)

		# get reward from bandit
		r = _bandit(a)
		_nu[a] = _nu[a] + 1
		_qu[a] = _qu[a] + (r - _qu[a]) / _nu[a]

		# Reduce tendency to explore with number of steps (or with age for humans)
		_epsilon = _epsilon * (1 - _rho)

		# Statistics
		_r_sum += r
		_n += 1
		_r_mean.append(_r_sum/_n)

	return np.array(_r_mean)


if __name__ == '__main__':
	# init bandits with different biases for shifting reward probability
	# -> expected reward q*(a)
	ql_star = np.array(np.linspace(-0.5, +0.5, K))

	# Init parameters
	num_realisations = 10
	episode_len = 10000

	# Init Q and N
	qu = np.zeros(K)
	nu = np.zeros(K)

	# Init r_mean
	r_mean = np.zeros(episode_len)
	for realization in range(0, num_realisations):
		# Re-Init Q and N
		qu = np.zeros(K)
		nu = np.zeros(K)
#		qu = np.random.normal(qu)
#		nu = np.random.normal(nu)
		r_mean += test(episode_len, qu, nu, _bandit=lambda a: bandit(a, ql_star))

		print(f"qu       = {qu}")
		print(f"nu       = {nu}")

	print(f"ql_star  = {ql_star}")

	pl.plot(r_mean/num_realisations)
	pl.grid()
	pl.title(f"Norm. E(R), num. realizations: Z={num_realisations}, num. bandits: K={K}")
	pl.xlabel("Episode")
	pl.ylabel("E(R)/Z")
	ax = pl.gca()
	ax.set_xlim([-episode_len/10, episode_len])
	ax.set_ylim([0, 1])

	pl.show()